Yan Deli, Senior Expert at Tencent
LLMs are probabilistic models; they do not understand causality.
Large language models (LLMs) do not truly understand language; they merely predict the next token (Next Token Prediction). They are essentially dull “word-chain” machines, proceeding one step at a time without retracing their path. At their core, LLMs are probabilistic models that learn, from statistical co-occurrence patterns, “what often appears alongside what,” producing outputs based on “what is typically the case.” They do not grasp “why things are this way” nor establish causal chains between variables. In other words, LLMs can recognize correlations but cannot comprehend causation.
LLMs primarily rely on probabilistic prediction—so why can they demonstrate such broad and deep capabilities? Why are they able to provide coherent causal explanations, even answering questions like “What would have happened if we had made a different choice?” This seems to go beyond the Association level of Pearl’s Causal Hierarchy and also encompasses the Intervention and Counterfactual levels. After all, compression is intelligence. The essence of large models is to compress vast amounts of knowledge into a limited set of connections (Geoffrey Hinton, 2025), and probabilistic data models and lossless compression are fundamentally equivalent. Language itself is a compression algorithm: humans encode extensive causal knowledge within text, and linguistic structures inherently carry interpretations of the world.
Since it is a probabilistic model, reliability issues are inherent. Hallucination—such as answering irrelevant questions, misattributing information, fabricating facts, citing incorrect sources, presenting contradictory statements, or demonstrating a lack of common sense—is unavoidable. These hallucinations stem from the technical characteristics of LLMs and are baked into their design; they can only be mitigated, not entirely eliminated. Even the models with the lowest global hallucination rates still exhibit a rate of 1.8%, as shown in the figure below.
Figure: Anchor hallucination rates of 25 mainstream large language models worldwide (May 2026)
LLMs have blind spots, lack insights, and cannot learn continuously.
LLMs are powerful, but they also have blind spots: they cannot learn tacit knowledge or sensitive information. Polanyi’s paradox states, “We can know more than we can tell.” Many tasks can be accomplished through intuition alone, yet their rules and steps cannot be articulated in language. For example, maintaining balance while cycling, gauging the appropriate level of empathy, the gentle touch of a fingertip, or a knowing glance—these kinds of knowledge that are “hard to articulate clearly and difficult to transfer effectively” are known as tacit knowledge. By contrast, there is explicit knowledge—knowledge that can be documented and digitized—and this is what LLMs can learn. Of course, not all explicit knowledge is accessible; its availability matters. Sensitive information and offline data are often more valuable, and a single piece of such information can be worth far more than ten thousand words—yet these remain blind spots for LLMs.
LLMs seem to know everything, yet they lack creative, disruptive insights (CDIs). Richard Sutton, a Turing Award winner and the “father of reinforcement learning,” pointed out in 2026: “These systems are essentially harnessing the power of human knowledge and delivering it back to us; they do not have the capacity to generate their own knowledge.” To assess whether LLMs can independently “rediscover” established scientific breakthroughs, Demis Hassabis introduced the Einstein Test this February. By restricting the knowledge base to before 1911, if an AI can independently derive Einstein’s 1915 theory of general relativity, it passes the Einstein Test. As Hassabis stated: “Clearly, current systems are still far from achieving this.”
Continuous learning is another critical missing piece. According to the six‑step roadmap toward AGI, continuous learning represents the next technological milestone to overcome after the current agent stage, and it is an essential phase on the path to self‑iterative systems (the singularity) and embodied intelligence. Continuous learning refers to the ability to keep acquiring new knowledge without forgetting what has already been learned—a capability that humans are born with. Unfortunately, LLMs do not possess this trait. LLMs are “frozen geniuses”: they rely on static training datasets, and once training is complete and deployment begins, their weights become fixed, leaving them unable to learn and grow from real‑world experience. The core challenge of continuous learning stems from the “stability–plasticity dilemma,” which entails two major obstacles: catastrophic forgetting (learning something new while losing previously acquired knowledge) and a loss of plasticity (a gradual decline in the ability to learn new tasks).
World Model: From Language to Action
“The limits of my language are the limits of my world.” “The world is everything that happens.” Wittgenstein’s assertion from a century ago has today revealed a remarkable power of prophetic insight.
Language is humanity’s most powerful invention, yet it falls short of fully capturing our world. The real world is not composed of words and sentences. LLMs are “wordsmiths in the dark”—eloquent and articulate, but devoid of any embodied experience in the physical world. Simply scaling language is no longer sufficient; we need to strengthen their intrinsic representations of the physical world, spatial structures, temporal continuity, and the consequences of actions. As Sutton puts it, LLMs are merely a “fleeting fascination,” capturing only a small fraction of true intelligence. Genuine intelligence must possess a model of the world.
A world model is a predictor of state changes, describing “how the environment’s state evolves over time and in response to actions.” At its core, it is a dynamical model, which can be represented by the following function:
A world model is a function that, given the current state st and the intended action at, predicts the next state ŝₜ₊₁ under parameters θ. Since the physical world is stochastic and partially observable, the output is typically a probability distribution over possible next states, which more closely reflects the real environment. The goal of a world model is not to predict a single, exact future, but rather to construct a probabilistic distribution of future outcomes that can be used for planning. World models “compute the uncomputable,” seeking those intangible, instinctive, fundamental insights into how the world operates. Fei-Fei Li and Yann LeCun are two of the most influential leaders in AI. Li created the ImageNet dataset and is revered as the “Godmother of AI”; LeCun received the 2018 Turing Award and is hailed as the “Father of Convolutional Neural Networks.” Both are also among the strongest advocates and supporters of world models. In 2026, Li categorized world models into three types: renderers that output pixel-level representations, simulators that generate state transitions, and planners that produce actionable policies. Her space‑intelligence startup, World Labs, was acquired by AMD for $8.2 billion. After leaving Meta, LeCun founded AMI Labs in his hometown of Paris at the end of 2025, with the company now valued at $3.5 billion.
World models have become a hot area for startups, with many companies claiming to be world-model companies. Industry giants such as NVIDIA and Google have also entered the field. NVIDIA has launched Cosmos, a foundational world-model platform, designed to help robots, autonomous vehicles, and other physical systems better understand and predict real-world environments; Google DeepMind positions Genie 3 as “a new frontier for world models.”
We should pay close attention to the “master entrepreneurship” phenomenon.
Major scientific discoveries and technological advances are often driven by young people. As Albert Einstein once said, “If a person has not made significant scientific contributions by the age of 30, they probably never will.” Paul Dirac, a key founder of quantum mechanics, also observed that most great physicists had already accomplished their most outstanding work before turning 30. In his book The Structure of Scientific Revolutions, Thomas Kuhn pointed out: “Almost all those who make fundamental breakthroughs that give rise to a new paradigm are either very young or have just entered the field they are about to transform.” This is known as the “30-year rule” of scientific and technological innovation—more accurately, the “newcomer rule.”
“Artificial intelligence is a young field, and it’s also a field for young people.” At age 24, Turing proposed the concept of the Turing machine, and at 38 he introduced the Turing test thought experiment; John McCarthy coined the term “Artificial Intelligence” at 28; the eight authors of “Attention Is All You Need” are all young as well; prodigy high school student Chen Guangyu joined Moonshot AI as a machine learning researcher at 16 and co-authored the paper “Attention Residuals”; Sam Altman dropped out of college at 19 to start a business, and at 30 co-founded OpenAI; another 19-year-old dropout was Wang Tao, who founded the data-labeling company Scale AI and joined Meta as Chief AI Officer at 28; Demis Hassabis left his job at 22 to start a venture, and at 34 founded DeepMind; Dario Amodei, at 38, founded Anthropic with six former OpenAI colleagues—already considered relatively advanced in age.
Youth doesn’t just refer to age; it also means they are “ordinary people” without any aura. For example, McCarthy was still an assistant professor when he organized the Dartmouth Conference; the “Transformer eight siblings” held relatively low-ranking positions and were little known when they published “Attention Is All You Need.”
However, the world of AI models is different: many accomplished scholars and professors have ventured into entrepreneurship. Leaders such as Yann LeCun and Fei-Fei Li were already 65 and 48 years old, respectively, when they founded their companies. In addition, Professor Yi Ma of the University of Hong Kong established Yisheng Technology at age 51, while Professor Anima Anandkumar of Caltech launched Accelerated Understanding at age 44; Professor Jun Zhu of Tsinghua University cofounded Shengshu Technology at age 40. Assistant Professor Biwei Huang of the University of California, San Diego, and Associate Professor Shuo Feng of Tsinghua University, though relatively younger, also started their companies around age 35. This is significantly higher than the average age of 29 for AI unicorn founders at the time of founding (source: Antler, 2026), as shown in the table below.
Table of some world-modeling companies founded by the professor
Entrepreneurship means breaking with convention and forging a new path, but it is also a high‑risk endeavor. Young people are bold and less constrained by outdated paradigms; by diving in and learning as they go, they are more likely to blaze a trail. By contrast, erudite, high‑profile professors often carry the burden of their public image and are locked into established ways of thinking, tending to be risk‑averse—hence, truly successful entrepreneurial ventures led by them are relatively rare. Today, as academic heavyweights flock to the world‑model arena, how should we interpret this trend? Setting aside cases where professors serve in name only while students do the real work, there are essentially two explanations.
First is the special theory of world models. In highly knowledge-intensive fields, it requires professors who possess deep expertise to found great companies.
Second is the AI bubble theory. In a bubble, investors are generous and willing to pay for all kinds of visions. The concept of a world model is exhilarating, and its fervor is a microcosm of the AI bubble.